def test_make_vec_env(self):
venv1 = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 5, 1,
True)
venv2 = make_vec_env(ParallelVecEnv, make_gym_env, 'CartPole-v1', 5, 1,
True)
assert isinstance(venv1, VecEnv) and isinstance(venv1, SerialVecEnv)
assert isinstance(venv2, VecEnv) and isinstance(venv2, ParallelVecEnv)
assert venv1.num_env == venv2.num_env
env_spec1 = EnvSpec(venv1)
assert env_spec1.num_env == venv1.num_env
env_spec2 = EnvSpec(venv2)
assert env_spec2.num_env == venv2.num_env
assert venv1.observation_space == venv2.observation_space
assert venv1.action_space == venv2.action_space
assert venv1.reward_range == venv2.reward_range
assert venv1.T == venv2.T
o1 = venv1.reset()
o2 = venv2.reset()
# Two environments should have same random seeds, then same results under same actions
assert np.allclose(o1, o2)
a = [1] * 5
o1, r1, d1, _ = venv1.step(a)
o2, r2, d2, _ = venv2.step(a)
assert np.allclose(o1, o2)
assert np.allclose(r1, r2)
assert np.allclose(d1, d2)
assert not venv1.closed
venv1.close()
assert venv1.closed
assert not venv2.closed
venv2.close()
assert venv2.closed
def test_make_gym_env(self):
env = make_gym_env(env_id='CartPole-v1', seed=0, monitor=False)
assert isinstance(env, Env)
assert not isinstance(env, gym.Env)
assert isinstance(env, Wrapper)
assert isinstance(env.observation_space, Box)
assert isinstance(env.action_space, Discrete)
env_spec = EnvSpec(env)
assert env_spec.control_type == 'Discrete'
assert env_spec.T == 500
assert env_spec.max_episode_reward == 475.0
assert env_spec.reward_range == (-float('inf'), float('inf'))
assert not env_spec.is_vec_env
with pytest.raises(TypeError):
env_spec.num_env
assert env.reset().shape == (4, )
assert len(env.step(env.action_space.sample())) == 4
del env
del env_spec
# Pendulum, continuous
# do not test redundant part
env = make_gym_env('Pendulum-v0', seed=0)
assert isinstance(env, Env)
env_spec = EnvSpec(env)
assert isinstance(env_spec.action_space, Box)
assert env_spec.T == 200
assert env_spec.control_type == 'Continuous'
assert env.reset().shape == (3, )
assert len(env.step(env.action_space.sample())) == 4
def test_random_agent():
env = make_gym_env('Pendulum-v0', 0)
env_spec = EnvSpec(env)
agent = RandomAgent(None, env_spec)
out = agent.choose_action(env.reset())
assert isinstance(out, dict)
assert 'action' in out and out['action'].shape == (1,)
venv = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(venv)
agent = RandomAgent(None, env_spec)
out = agent.choose_action(env.reset())
assert isinstance(out, dict)
assert 'action' in out and len(out['action']) == 3 and isinstance(out['action'][0], int)
def test_random_policy():
env = make_gym_env('Pendulum-v0', 0)
env_spec = EnvSpec(env)
policy = RandomPolicy(None, env_spec)
out = policy(env.reset())
assert isinstance(out, dict)
assert 'action' in out and out['action'].shape == (1, )
venv = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v0', 3, 0, False)
env_spec = EnvSpec(venv)
policy = RandomPolicy(None, env_spec)
out = policy(env.reset())
assert isinstance(out, dict)
assert 'action' in out and len(out['action']) == 3 and isinstance(
out['action'][0], int)
def test_terminal_state_from_episode():
env = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(env)
with pytest.raises(AssertionError):
terminal_state_from_episode([1, 2, 3])
D = BatchEpisode(env_spec)
D.obs[0] = [0.1, 0.2, 1.3]
D.done[0] = [False, False, True]
D.info[0] = [{}, {}, {'terminal_observation': 0.3}]
D.obs[1] = [1, 2, 3, 4, 5, 6, 7, 8, 9]
D.done[1] = [False] * 9
D.obs[2] = [10, 15]
D.done[2] = [False, True]
D.info[2] = [{}, {'terminal_observation': 20}]
terminal_states = terminal_state_from_episode(D)
assert terminal_states.shape == (2, ) + env_spec.observation_space.shape
assert np.allclose(terminal_states[0], 0.3)
assert np.allclose(terminal_states[1], 20)
D.done[0][-1] = False
D.done[2][-1] = False
assert terminal_state_from_episode(D) is None
with pytest.raises(AssertionError):
terminal_state_from_segment(D)
def test_returns_from_episode():
env = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(env)
D = BatchEpisode(env_spec)
D.r[0] = [1, 2, 3]
D.done[0] = [False, False, True]
D.r[1] = [1, 2, 3, 4, 5]
D.done[1] = [False, False, False, False, False]
D.r[2] = [1, 2, 3, 4, 5, 6, 7, 8]
D.done[2] = [False, False, False, False, False, False, False, True]
out = returns_from_episode(D, 1.0)
assert out.shape == (3, D.maxT)
assert np.allclose(out[0], [6, 5, 3, 0, 0, 0, 0, 0])
assert np.allclose(out[1], [15, 14, 12, 9, 5, 0, 0, 0])
assert np.allclose(out[2], [36, 35, 33, 30, 26, 21, 15, 8])
del out
out = returns_from_episode(D, 0.1)
assert out.shape == (3, D.maxT)
assert np.allclose(out[0], [1.23, 2.3, 3, 0, 0, 0, 0, 0])
assert np.allclose(out[1], [1.2345, 2.345, 3.45, 4.5, 5, 0, 0, 0])
assert np.allclose(
out[2], [1.2345678, 2.345678, 3.45678, 4.5678, 5.678, 6.78, 7.8, 8])
with pytest.raises(AssertionError):
returns_from_segment(D, 0.1)
def test_segment_runner(env_id):
env = make_vec_env(SerialVecEnv, make_gym_env, env_id, 3, 0, True)
env_spec = EnvSpec(env)
agent = RandomAgent(None, env_spec)
runner = SegmentRunner(None, agent, env)
D = runner(4)
assert len(D) == 3
assert all([isinstance(d, Segment) for d in D])
assert all([d.T == 4 for d in D])
# Check if s in transition is equal to s_next in previous transition
for d in D:
for s1, s2 in zip(d.transitions[:-1], d.transitions[1:]):
assert np.allclose(s1.s_next, s2.s)
# Long horizon
D = runner(T=1000)
for d in D:
assert d.T == 1000
with pytest.raises(AssertionError):
env = make_vec_env(SerialVecEnv, make_gym_env, env_id, 3, 0, False)
SegmentRunner(None, agent, env)
def test_bootstrapped_returns_from_segment():
env = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(env)
D = BatchSegment(env_spec, 5)
D.r[0] = [1, 2, 3, 4, 5]
D.done[0] = [False, False, False, False, False]
D.r[1] = [1, 2, 3, 4, 5]
D.done[1] = [False, False, True, False, False]
D.r[2] = [1, 2, 3, 4, 5]
D.done[2] = [True, False, False, False, True]
last_Vs = torch.tensor([10, 20, 30]).unsqueeze(1)
out = bootstrapped_returns_from_segment(D, last_Vs, 1.0)
assert out.shape == (3, 5)
assert np.allclose(out[0], [25, 24, 22, 19, 15])
assert np.allclose(out[1], [6, 5, 3, 29, 25])
assert np.allclose(out[2], [1, 14, 12, 9, 5])
del out
out = bootstrapped_returns_from_segment(D, last_Vs, 0.1)
assert out.shape == (3, 5)
assert np.allclose(out[0], [1.2346, 2.346, 3.46, 4.6, 6])
assert np.allclose(out[1], [1.23, 2.3, 3, 4.7, 7])
assert np.allclose(out[2], [1, 2.345, 3.45, 4.5, 5])
with pytest.raises(AssertionError):
bootstrapped_returns_from_episode(D, last_Vs, 0.1)
def test_returns_from_segment():
env = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(env)
D = BatchSegment(env_spec, 5)
D.r[0] = [1, 2, 3, 4, 5]
D.done[0] = [False, False, False, False, False]
D.r[1] = [1, 2, 3, 4, 5]
D.done[1] = [False, False, True, False, False]
D.r[2] = [1, 2, 3, 4, 5]
D.done[2] = [True, False, False, False, True]
out = returns_from_segment(D, 1.0)
assert out.shape == (3, 5)
assert np.allclose(out[0], [15, 14, 12, 9, 5])
assert np.allclose(out[1], [6, 5, 3, 9, 5])
assert np.allclose(out[2], [1, 14, 12, 9, 5])
del out
out = returns_from_segment(D, 0.1)
assert out.shape == (3, 5)
assert np.allclose(out[0], [1.2345, 2.345, 3.45, 4.5, 5])
assert np.allclose(out[1], [1.23, 2.3, 3, 4.5, 5])
assert np.allclose(out[2], [1, 2.345, 3.45, 4.5, 5])
with pytest.raises(AssertionError):
returns_from_episode(D, 0.1)
def test_bootstrapped_returns_from_episode():
env = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(env)
D = BatchEpisode(env_spec)
D.r[0] = [1, 2, 3]
D.done[0] = [False, False, True]
D.completed[0] = True
D.r[1] = [1, 2, 3, 4, 5]
D.done[1] = [False, False, False, False, False]
D.completed[1] = False
D.r[2] = [1, 2, 3, 4, 5, 6, 7, 8]
D.done[2] = [False, False, False, False, False, False, False, True]
D.completed[2] = True
last_Vs = torch.tensor([10, 20, 30]).unsqueeze(1)
out = bootstrapped_returns_from_episode(D, last_Vs, 1.0)
assert out.shape == (3, D.maxT)
assert np.allclose(out[0], [6, 5, 3, 0, 0, 0, 0, 0])
assert np.allclose(out[1], [35, 34, 32, 29, 25, 0, 0, 0])
assert np.allclose(out[2], [36, 35, 33, 30, 26, 21, 15, 8])
del out
out = bootstrapped_returns_from_episode(D, last_Vs, 0.1)
assert out.shape == (3, D.maxT)
assert np.allclose(out[0], [1.23, 2.3, 3, 0, 0, 0, 0, 0])
assert np.allclose(out[1], [1.2347, 2.347, 3.47, 4.7, 7, 0, 0, 0])
assert np.allclose(
out[2], [1.2345678, 2.345678, 3.45678, 4.5678, 5.678, 6.78, 7.8, 8])
with pytest.raises(AssertionError):
bootstrapped_returns_from_segment(D, last_Vs, 0.1)
def test_final_state_from_segment():
env = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(env)
with pytest.raises(AssertionError):
final_state_from_segment([1, 2, 3])
D = BatchSegment(env_spec, 4)
D.obs = np.random.randn(*D.obs.shape)
D.done.fill(False)
D.done[0, -1] = True
D.info[0] = [{}, {}, {}, {'terminal_observation': [0.1, 0.2, 0.3, 0.4]}]
D.done[1, 2] = True
D.info[1] = [{}, {}, {'terminal_observation': [1, 2, 3, 4]}, {}]
D.done[2, -1] = True
D.info[2] = [{}, {}, {}, {'terminal_observation': [10, 20, 30, 40]}]
final_states = final_state_from_segment(D)
assert final_states.shape == (3, ) + env_spec.observation_space.shape
assert np.allclose(final_states[0], [0.1, 0.2, 0.3, 0.4])
assert np.allclose(final_states[1], D.numpy_observations[1, -1, ...])
assert not np.allclose(final_states[1], [1, 2, 3, 4])
assert np.allclose(final_states[2], [10, 20, 30, 40])
with pytest.raises(AssertionError):
final_state_from_episode(D)
def __init__(self, config, agent, env):
super().__init__(config, agent, env)
self.env_spec = EnvSpec(self.env)
self.obs_buffer = None # for next call
self.done_buffer = None # masking
def test_sticky_agent():
sticky_action = 0
env = make_gym_env('CartPole-v1', 0)
env_spec = EnvSpec(env)
agent = StickyAgent(None, env_spec, sticky_action)
out = agent.choose_action(env.reset())
assert isinstance(out, dict)
assert 'action' in out and isinstance(out['action'], int)
assert out['action'] == sticky_action
venv = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(venv)
agent = StickyAgent(None, env_spec, sticky_action)
out = agent.choose_action(env.reset())
assert isinstance(out, dict)
assert 'action' in out and len(out['action']) == 3 and isinstance(out['action'][0], int)
assert np.allclose(out['action'], [0, 0, 0])
def test_categorical_head():
with pytest.raises(AssertionError):
env = make_gym_env('Pendulum-v0', 0)
env_spec = EnvSpec(env)
CategoricalHead(None, None, 30, env_spec)
env = make_gym_env('CartPole-v1', 0)
env_spec = EnvSpec(env)
head = CategoricalHead(None, None, 30, env_spec)
assert head.feature_dim == 30
assert isinstance(head.action_head, nn.Linear)
assert head.action_head.in_features == 30 and head.action_head.out_features == 2
dist = head(torch.randn(3, 30))
assert isinstance(dist, Categorical)
assert list(dist.batch_shape) == [3]
assert list(dist.probs.shape) == [3, 2]
action = dist.sample()
assert action.shape == (3, )
def check(agent_name, env_name):
# Create environment
list_make_env = make_envs(make_env=make_gym_env,
env_id=env_name,
num_env=2,
init_seed=0)
env = SerialVecEnv(list_make_env=list_make_env)
env_spec = EnvSpec(env)
assert env.num_env == 2
# Create agent
if agent_name == 'random':
agent = RandomAgent(env_spec=env_spec, config=None)
elif agent_name == 'agent1':
agent = Agent1(config=None)
elif agent_name == 'agent2':
agent = Agent2(config=None)
else:
raise ValueError('Wrong agent name')
# Create runner
runner = SegmentRunner(agent=agent, env=env, gamma=1.0)
# Small batch
D = runner(T=3, reset=False)
assert len(D) == 2
assert all([isinstance(d, Segment) for d in D])
assert all([d.T == 3 for d in D])
assert all([d.gamma == 1.0 for d in D])
# Check additional information
for d in D:
for t in d.transitions:
if agent_name != 'random':
assert 'action_logprob' in t.info
# Check if s in transition is equal to s_next in previous transition
for d in D:
for s1, s2 in zip(d.transitions[:-1], d.transitions[1:]):
assert np.allclose(s1.s_next, s2.s)
# Take one more step, test rolling effect, i.e. first state should be same as last state in previous D
D2 = runner(T=1, reset=False)
assert len(D2) == 2
assert all([d.T == 1 for d in D2])
for d, d2 in zip(D, D2):
assert np.allclose(d2.all_s[0][0], d.transitions[-1].s_next)
# Long horizon
D = runner(T=200, reset=True)
# Segment with identical time steps
assert all([d.T == 200 for d in D])
# For CartPole, 200 time steps, should be somewhere done=True
if env_name == 'CartPole-v1':
assert any([True in d.all_done for d in D])
assert all([len(d.trajectories) > 1 for d in D])
def make_env_spec(self):
list_make_env = make_envs(make_env=make_gym_env,
env_id='CartPole-v1',
num_env=3,
init_seed=0)
venv = SerialVecEnv(list_make_env=list_make_env, rolling=True)
env_spec = EnvSpec(venv)
return env_spec
def make_env_spec(self):
list_make_env = make_envs(make_env=make_gym_env,
env_id='Pendulum-v0',
num_env=1,
init_seed=0)
venv = SerialVecEnv(list_make_env=list_make_env)
env_spec = EnvSpec(venv)
return env_spec
def check(agent_name, env_name):
# Create environment
list_make_env = make_envs(make_env=make_gym_env,
env_id=env_name,
num_env=1,
init_seed=0)
env = SerialVecEnv(list_make_env=list_make_env)
env_spec = EnvSpec(env)
# Create agent
if agent_name == 'random':
agent = RandomAgent(env_spec=env_spec, config=None)
elif agent_name == 'agent1':
agent = Agent1(config=None)
elif agent_name == 'agent2':
agent = Agent2(config=None)
else:
raise ValueError('Wrong agent name')
# Test: not allowed more than one environment for TrajectoryRunner
with pytest.raises(AssertionError):
list_make_env2 = make_envs(make_env=make_gym_env,
env_id=env_name,
num_env=2,
init_seed=0)
env2 = SerialVecEnv(list_make_env=list_make_env2)
runner2 = TrajectoryRunner(agent=agent, env=env2, gamma=1.0)
# Create runner
runner = TrajectoryRunner(agent=agent, env=env, gamma=1.0)
# Small batch
D = runner(N=3, T=4)
assert len(D) == 3
assert all([isinstance(d, Trajectory) for d in D])
assert all([d.T == 4 for d in D])
assert all([d.gamma == 1.0 for d in D])
# Check additional information
for d in D:
for t in d.transitions:
if agent_name != 'random':
assert 'action_logprob' in t.info
# Check if s in transition is equal to s_next in previous transition
for d in D:
for s1, s2 in zip(d.transitions[:-1], d.transitions[1:]):
assert np.allclose(s1.s_next, s2.s)
# Long horizon
D = runner(N=3, T=1000)
for d in D:
if d.T < 1000:
assert d.all_done[-1] == True
def test_diag_gaussian_head():
with pytest.raises(AssertionError):
env = make_gym_env('CartPole-v1', 0)
env_spec = EnvSpec(env)
DiagGaussianHead(None, None, 30, env_spec)
env = make_gym_env('Pendulum-v0', 0)
env_spec = EnvSpec(env)
head = DiagGaussianHead(None, None, 30, env_spec)
assert head.feature_dim == 30
assert isinstance(head.mean_head, nn.Linear)
assert isinstance(head.logstd_head, nn.Parameter)
assert head.mean_head.in_features == 30 and head.mean_head.out_features == 1
assert list(head.logstd_head.shape) == [1]
assert torch.eq(head.logstd_head, torch.tensor(-0.510825624))
dist = head(torch.randn(3, 30))
assert isinstance(dist, Independent) and isinstance(dist.base_dist, Normal)
assert list(dist.batch_shape) == [3]
action = dist.sample()
assert list(action.shape) == [3, 1]
head = DiagGaussianHead(None, None, 30, env_spec, std_style='softplus')
dist = head(torch.randn(3, 30))
action = dist.sample()
assert list(action.shape) == [3, 1]
assert torch.eq(head.logstd_head, torch.tensor(-0.19587036834631966))
head = DiagGaussianHead(None, None, 30, env_spec, std_style='sigmoidal')
assert torch.eq(head.logstd_head, torch.tensor(-0.871222446472449))
head = DiagGaussianHead(None, None, 30, env_spec, std_state_dependent=True)
dist = head(torch.randn(3, 30))
action = dist.sample()
assert list(action.shape) == [3, 1]
head = DiagGaussianHead(None, None, 30, env_spec, constant_std=0.3)
dist = head(torch.randn(3, 30))
action = dist.sample()
assert list(action.shape) == [3, 1]
assert not head.logstd_head.requires_grad
assert torch.eq(head.logstd_head, torch.tensor([-1.2039728]))
def __call__(self, config):
# Set random seeds: PyTorch, numpy.random, random
set_global_seeds(seed=config['seed'])
# Create environment and seed it
env = make_env(seed=config['seed'],
monitor=False,
monitor_dir=None)
# Create environment specification
env_spec = EnvSpec(env) # TODO: integrate within make_env globally
# Create device
device = torch.device('cuda' if config['cuda'] else 'cpu')
# Create logger
logger = Logger(name='logger')
# Create policy
network = MLP(config=config)
policy = CategoricalPolicy(network=network, env_spec=env_spec)
policy.network = policy.network.to(device)
# Create optimizer
optimizer = optim.Adam(policy.network.parameters(), lr=config['lr'])
# Learning rate scheduler
max_epoch = config['train_iter'] # Max number of lr decay, Note where lr_scheduler put
lambda_f = lambda epoch: 1 - epoch/max_epoch # decay learning rate for each training epoch
lr_scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda_f)
# Create agent
agent_class = ActorCriticAgent#REINFORCEAgent
agent = agent_class(policy=policy,
optimizer=optimizer,
config=config,
lr_scheduler=lr_scheduler,
device=device)
# Create runner
runner = Runner(agent=agent,
env=env,
gamma=config['gamma'])
# Create engine
engine = Engine(agent=agent,
runner=runner,
config=config,
logger=logger)
# Training
train_output = engine.train()
np.save('logs/returns_ActorCritic', train_output)
return None
def test_constraint_action():
env = make_gym_env('Pendulum-v0', 0)
env_spec = EnvSpec(env)
action = torch.tensor([1.5])
assert torch.eq(constraint_action(env_spec, action), torch.tensor([1.5]))
action = torch.tensor([3.0])
assert torch.eq(constraint_action(env_spec, action), torch.tensor([2.0]))
action = torch.tensor([-10.0])
assert torch.eq(constraint_action(env_spec, action), torch.tensor([-2.0]))
def test_env_spec(self):
env = gym.make('CartPole-v1')
env = GymWrapper(env)
env.seed(0)
env_spec = EnvSpec(env)
assert isinstance(env_spec.observation_space, Box)
assert isinstance(env_spec.action_space, Discrete)
assert env_spec.control_type == 'Discrete'
assert env_spec.T == 500
assert env_spec.max_episode_reward == 475.0
assert env_spec.reward_range == (-float('inf'), float('inf'))
def test_gae_from_episode():
env = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(env)
D = BatchEpisode(env_spec)
D.r[0] = [1, 2, 3]
D.done[0] = [False, False, True]
D.completed[0] = True
D.r[1] = [1, 2, 3, 4, 5]
D.done[1] = [False, False, False, False, False]
D.completed[1] = False
D.r[2] = [1, 2, 3, 4, 5, 6, 7, 8]
D.done[2] = [False, False, False, False, False, False, False, True]
D.completed[2] = True
all_Vs = [
torch.tensor([[0.1], [0.5], [1.0]]),
torch.tensor([[1.1], [1.5], [2.0]]),
torch.tensor([[2.1], [2.5], [3.0]]),
torch.tensor([[3.1], [3.5], [4.0]]),
torch.tensor([[4.1], [4.5], [5.0]]),
torch.tensor([[5.1], [5.5], [6.0]]),
torch.tensor([[6.1], [6.5], [7.0]]),
torch.tensor([[7.1], [7.5], [8.0]])
]
last_Vs = torch.tensor([10, 20, 30]).unsqueeze(1)
all_Vs = torch.stack(all_Vs, 1)
out = gae_from_episode(D, all_Vs, last_Vs, 1.0, 0.5)
assert out.shape == (3, D.maxT)
assert np.allclose(out[0], [3.725, 3.45, 0.9, 0, 0, 0, 0, 0])
assert np.allclose(out[1], [6.40625, 8.8125, 11.625, 15.25, 20.5, 0, 0, 0])
assert np.allclose(out[2],
[5.84375, 7.6875, 9.375, 10.75, 11.5, 11., 8, 0.])
del out
out = gae_from_episode(D, all_Vs, last_Vs, 0.1, 0.2)
assert out.shape == (3, D.maxT)
assert np.allclose(out[0], [1.03256, 1.128, 0.9, 0, 0, 0, 0, 0])
assert np.allclose(out[1], [0.665348, 0.7674, 0.87, 1, 2.5, 0, 0, 0])
assert np.allclose(out[2], [
0.206164098, 0.308204915, 0.410245728, 0.5122864, 0.61432, 0.716, 0.8,
0
])
with pytest.raises(AssertionError):
gae_from_segment(D, all_Vs, last_Vs, 0.1, 0.2)
def test_episode_runner(vec_env, env_id):
env = make_vec_env(vec_env, make_gym_env, env_id, 3, 0)
env_spec = EnvSpec(env)
if env_id == 'CartPole-v1':
sticky_action = 1
elif env_id == 'Pendulum-v0':
sticky_action = [0.1]
T = 30
agent = StickyAgent(None, env_spec, sticky_action)
runner = EpisodeRunner(None, agent, env)
D = runner(T)
assert D.N == 3
assert D.maxT == max(D.Ts)
seeder = Seeder(0)
seed1, seed2, seed3 = seeder(3)
env1 = make_gym_env(env_id, seed1)
env2 = make_gym_env(env_id, seed2)
env3 = make_gym_env(env_id, seed3)
for n, ev in enumerate([env1, env2, env3]):
obs = ev.reset()
assert np.allclose(obs, D.observations[n][0])
assert np.allclose(obs, D.numpy_observations[n, 0, ...])
for t in range(T):
obs, reward, done, info = ev.step(sticky_action)
assert np.allclose(reward, D.rewards[n][t])
assert np.allclose(reward, D.numpy_rewards[n, t])
assert np.allclose(done, D.dones[n][t])
assert done == D.numpy_dones[n, t]
assert int(not done) == D.masks[n][t]
assert int(not done) == D.numpy_masks[n, t]
if done:
assert np.allclose(obs, D.infos[n][t]['terminal_observation'])
assert D.completes[n]
assert np.allclose(0.0, D.numpy_observations[n, t + 1 + 1:,
...])
assert np.allclose(0.0, D.numpy_actions[n, t + 1:, ...])
assert np.allclose(0.0, D.numpy_rewards[n, t + 1:])
assert np.allclose(True, D.numpy_dones[n, t + 1:])
assert np.allclose(0.0, D.numpy_masks[n, t + 1:])
break
else:
assert np.allclose(obs, D.observations[n][t + 1])
def test_td0_error_from_episode():
env = make_vec_env(SerialVecEnv, make_gym_env, 'CartPole-v1', 3, 0)
env_spec = EnvSpec(env)
D = BatchEpisode(env_spec)
D.r[0] = [1, 2, 3]
D.done[0] = [False, False, True]
D.completed[0] = True
D.r[1] = [1, 2, 3, 4, 5]
D.done[1] = [False, False, False, False, False]
D.completed[1] = False
D.r[2] = [1, 2, 3, 4, 5, 6, 7, 8]
D.done[2] = [False, False, False, False, False, False, False, True]
D.completed[2] = True
all_Vs = [
torch.tensor([[0.1], [0.5], [1.0]]),
torch.tensor([[1.1], [1.5], [2.0]]),
torch.tensor([[2.1], [2.5], [3.0]]),
torch.tensor([[3.1], [3.5], [4.0]]),
torch.tensor([[4.1], [4.5], [5.0]]),
torch.tensor([[5.1], [5.5], [6.0]]),
torch.tensor([[6.1], [6.5], [7.0]]),
torch.tensor([[7.1], [7.5], [8.0]])
]
last_Vs = torch.tensor([10, 20, 30]).unsqueeze(1)
all_Vs = torch.stack(all_Vs, 1)
out = td0_error_from_episode(D, all_Vs, last_Vs, 1.0)
assert out.shape == (3, D.maxT)
assert np.allclose(out[0], [2.0, 3, 0.9, 0, 0, 0, 0, 0])
assert np.allclose(out[1], [2, 3, 4, 5, 20.5, 0, 0, 0])
assert np.allclose(out[2], [2, 3, 4, 5, 6, 7, 8, 0])
del out
out = td0_error_from_episode(D, all_Vs, last_Vs, 0.1)
assert out.shape == (3, D.maxT)
assert np.allclose(out[0], [1.01, 1.11, 0.9, 0, 0, 0, 0, 0])
assert np.allclose(out[1], [0.65, 0.75, 0.85, 0.95, 2.5, 0, 0, 0])
assert np.allclose(out[2], [0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0])
with pytest.raises(AssertionError):
td0_error_from_segment(D, all_Vs, last_Vs, 0.1)
def init(self, seed, config):
# Make environment
# Remember to seed it in each working function !
self.env = make_vec_env(vec_env_class=SerialVecEnv,
make_env=make_gym_env,
env_id=config['env.id'],
num_env=1,
init_seed=seed)
self.env_spec = EnvSpec(self.env)
# Make agent
self.network = Network(config=config, env_spec=self.env_spec)
if self.env_spec.control_type == 'Discrete':
self.policy = CategoricalPolicy(config=config, network=self.network, env_spec=self.env_spec)
elif self.env_spec.control_type == 'Continuous':
self.policy = GaussianPolicy(config=config, network=self.network, env_spec=self.env_spec)
self.agent = Agent(policy=self.policy, config=config)
def _prepare(self, config):
self.env = make_vec_env(SerialVecEnv, make_gym_env, config['env.id'],
config['train.N'], 0)
self.env = VecClipAction(self.env)
if config['env.standardize']:
self.env = VecStandardize(self.env,
use_obs=True,
use_reward=False,
clip_obs=10.0,
clip_reward=10.0,
gamma=0.99,
eps=1e-08)
self.env_spec = EnvSpec(self.env)
self.device = torch.device('cpu')
self.agent = Agent(config, self.env_spec, self.device)
def test_vec_monitor(env_id, vec_env_class, num_env, seed, runner_class):
env = make_vec_env(vec_env_class, make_gym_env, env_id, num_env, seed)
env_spec = EnvSpec(env)
env = VecMonitor(env)
agent = RandomAgent(None, env_spec)
runner = runner_class(None, agent, env)
D = runner(1050)
for infos in D.infos:
for info in infos:
if 'terminal_observation' in info:
assert 'episode' in info
assert 'return' in info['episode'] and isinstance(
info['episode']['return'], np.float32)
assert 'horizon' in info['episode'] and isinstance(
info['episode']['horizon'], np.int32)
assert 'time' in info['episode'] and isinstance(
info['episode']['time'], float)
def test_rolling_segment_runner(vec_env, env_id):
env = make_vec_env(vec_env, make_gym_env, env_id, 3, 0)
env_spec = EnvSpec(env)
if env_id == 'CartPole-v1':
sticky_action = 1
elif env_id == 'Pendulum-v0':
sticky_action = [0.1]
T = 30
agent = StickyAgent(None, env_spec, sticky_action)
runner = RollingSegmentRunner(None, agent, env)
D = runner(T)
assert D.N == 3
assert D.T == T
seeder = Seeder(0)
seed1, seed2, seed3 = seeder(3)
env1 = make_gym_env(env_id, seed1)
env2 = make_gym_env(env_id, seed2)
env3 = make_gym_env(env_id, seed3)
for n, ev in enumerate([env1, env2, env3]):
obs = ev.reset()
assert np.allclose(obs, D.numpy_observations[n, 0, ...])
for t in range(T):
obs, reward, done, info = ev.step(sticky_action)
if done:
info['terminal_observation'] = obs
obs = ev.reset()
assert np.allclose(obs, D.numpy_observations[n, t + 1, ...])
assert np.allclose(sticky_action, D.numpy_actions[n, t, ...])
assert np.allclose(reward, D.numpy_rewards[n, t])
assert done == D.numpy_dones[n, t]
assert int(not done) == D.numpy_masks[n, t]
if done:
assert np.allclose(info['terminal_observation'],
D.infos[n][t]['terminal_observation'])
def test_vec_env(self, vec_env_class):
# unpack class
v_id, vec_env_class = vec_env_class
venv = make_vec_env(vec_env_class, make_gym_env, 'CartPole-v1', 5, 1,
True)
assert isinstance(venv, VecEnv)
assert v_id in [0, 1]
if v_id == 0:
isinstance(venv, SerialVecEnv)
elif v_id == 1:
assert isinstance(venv, ParallelVecEnv)
assert venv.num_env == 5
assert not venv.closed and venv.viewer is None
assert venv.unwrapped is venv
assert isinstance(venv.observation_space, Box)
assert isinstance(venv.action_space, Discrete)
assert venv.T == 500
assert venv.max_episode_reward == 475.0
assert venv.reward_range == (-float('inf'), float('inf'))
obs = venv.reset()
assert len(obs) == 5
assert np.asarray(obs).shape == (5, 4)
assert all([not np.allclose(obs[0], obs[i]) for i in [1, 2, 3, 4]])
a = [1] * 5
obs, rewards, dones, infos = venv.step(a)
assert all([len(item) == 5 for item in [obs, rewards, dones, infos]])
assert all([not np.allclose(obs[0], obs[i]) for i in [1, 2, 3, 4]])
# EnvSpec
env_spec = EnvSpec(venv)
assert isinstance(env_spec.action_space, Discrete)
assert isinstance(env_spec.observation_space, Box)
assert env_spec.control_type == 'Discrete'
assert env_spec.T == 500
assert env_spec.max_episode_reward == 475.0
assert env_spec.reward_range == (-float('inf'), float('inf'))
assert env_spec.is_vec_env
venv.close()
assert venv.closed